An adaptive multi-rate wideband (AMR-WB) speech codec with a sampling rate of 16 kHz is known as one of the speech codecs employed in handheld devices that support 4G mobile communication systems. When applied to smartphones, it provides a superior speech quality relative to conventional speech codecs. Nonetheless, a major disadvantage is that an algebraic codebook search occupies a significant computational load in an AMR-WB encoder. In other words, the high computational complexity accounts for the high power consumption on a smartphone battery. This paper presents an improved version of depth-first tree search (DF) algorithm as a means to considerably reduce the complexity of an algebraic codebook search in an AMR-WB speech codec. This proposed search algorithm firstly involves the choice of a specified number of candidate pulses according to a pulse contribution ranking. Subsequently, a DF search is performed on the candidate pulses for a set of best pulses. Consequently, the target of the search and computational complexity reduction can be reached as expected. With a well maintained speech quality, this proposal demonstrates a search performance superiority over a DF and a global pulse replacement approach. Furthermore, with DF as a benchmark, a computational load reduction above 73% is reached in all coding modes.
- algebraic code-excited linear-prediction (ACELP)
- algebraic codebook search
- depth-first tree search
- speech codec